sherpa-onnx
k2-fsa/sherpa-onnxReviewed by AI
Checked by RepoPicks Radar on
What is sherpa-onnx?
sherpa-onnx is a speech toolkit that runs on your own machine instead of a cloud service. It can transcribe speech, read text out loud, clone a voice, tell speakers apart, detect when someone is talking, add punctuation, clean up noisy audio and separate voice from music. It is mostly for people who build programs: there are APIs for 12 languages, from C++ and Python to Java and Rust, plus Android, iOS and web builds. For a quick test there are ready Windows programs and browser demos, and you download the speech models separately.
sherpa-onnx is free and open source (Apache-2.0) and can replace ElevenLabs, Otter.ai, and Google Cloud Speech-to-Text.
How easy is it to install sherpa-onnx?
A · download and open · Web
- Windows B
- macOS C
- Linux C
- Web A
The easiest route is the browser: the README points to hosted demo pages where, in its words, "All you need is a browser". On Windows the releases include ready programs such as sherpa-onnx-non-streaming-tts-x64-v2.0.1.exe and sherpa-onnx-streaming-asr-x64-v2.0.1.exe (19-22 MB), where the only extra step is downloading a speech model, so B. On Mac and Linux the releases are code libraries instead of finished programs, so you install Python or another language and run commands, which is a C.
sherpa-onnx on GitHub
- Stars
- 15,866on Nov 10, 2026
- Downloads
- 38,716,831535,808 this week
- Growth
- +124stars in 6 days
Stars over time
Facts
- Runs on
- Windows, macOS, Linux, Web
- You need to install
- Nothing extra
- To build from source
- Python, NVIDIA graphics card
- Disk space
- 22 MB–66 MB · from the installer
- Accounts or keys
- None
- Cost
- Free · Apache-2.0 licence, no paid tier and no account.
- License
- Apache-2.0 (Permissive)
- Can replace
- ElevenLabs, Otter.ai, Cloud Speech-to-Text
- Last change on GitHub
- Nov 10, 2026
sherpa-onnx risk signals (2)
- Infovoice cloning included
The README lists demo pages for "Voice cloning with ZipVoice (Chinese+English)" and "Voice cloning with Pocket TTS (English)". Copying someone else's voice without their permission can break the law in some countries.
AI review
- Infomodels downloaded separately
The README has a "Links for pre-trained models" section pointing to outside download pages for speech recognition, text-to-speech, VAD, keyword spotting and speaker ID models. The program files are 22-66 MB, but the models you add are not counted in that.
AI review
Evidence is shown in its original language.
What Radar could not check about sherpa-onnx
- How good the transcription and the synthetic voices sound in your own language, since quality depends on which model you download
- How much disk space you end up using after downloading models, which is not part of the 22-66 MB program files
- Whether the ready Windows programs work without extra Windows components
- Whether every listed board, phone system and NPU really works as the tables claim
- Whether the hosted browser demos keep running, since they are on outside services